Google launches Gemini 4 Carbon for cybersecurity and coding

A specialized tool for complex tasks

Google has introduced Gemini 4 Carbon, an artificial intelligence model specifically engineered for high-level coding and cybersecurity work. Unlike standard AI chatbots that provide quick answers to general questions, this model is built to handle technical problems that require sustained effort. It is designed to manage multi-step processes that involve deep investigation into software systems.

The model is intended to assist cybersecurity professionals in identifying weaknesses that hackers might use to breach systems. It can also evaluate whether a specific vulnerability is actually exploitable by an outside party. Beyond just finding problems, Carbon can generate the specific code changes needed to repair the underlying issues in a project.

Google notes that the model has been observed working for days on complex engineering tasks. This persistence allows it to tackle difficult bugs in large software projects by investigating root causes. By focusing on these demanding technical tasks, the model aims to help teams protect digital systems more effectively.

Performance and benchmark results

To measure the model's effectiveness, Google conducted several benchmark evaluations without using standard guardrails. The model achieved a score of 88.2% on SEC-bench Pro, which is a standard test for identifying software vulnerabilities. It also scored 54.8% on ExploitGym, a benchmark used to evaluate how well an AI can exploit a vulnerability.

Additionally, the model received a 71% score on CWE-bench, which measures the ability of an AI to actually fix the vulnerabilities it finds. These tests provide a snapshot of the model's capabilities in a controlled environment. However, Google clarifies that these results do not guarantee the model will perform at the same rate on real-world software.

Internal testing also showed significant improvements over previous versions of the technology. Google reported that Gemini 4 Carbon more than doubled the success rate of secure-code generation when compared to its predecessor, Gemini 4 Argon. This suggests a major leap forward in the model's ability to produce safe and functional code for developers.

Controlled access and safety

Because a model capable of finding software flaws could be misused by malicious actors, Google is not releasing it to the general public immediately. The company is taking a gradual deployment approach to ensure that the technology is used safely. They are currently offering access to a select group of cybersecurity defenders first.

This initial group will provide the feedback necessary for Google to evaluate the model's safeguards. Once the company gathers enough data and refines the safety protocols, they plan to expand access to other trusted partners. Broader access for general consumers, businesses, and developers is expected in the future, though no specific date has been set.

The goal of this phased rollout is to balance the utility of the tool with the risks it poses. By starting with professionals who understand the security implications, Google can monitor how the model handles real-world engineering problems. This allows the company to refine the AI's behavior before it reaches a wider audience.

Pricing and technical details

For developers who gain access to the model via an application programming interface (API), Google has set introductory pricing. The cost is currently $2 per million input tokens and $10 per million output tokens. Tokens are the basic units of text or data that the AI processes or generates during a task.

To encourage efficiency, Google is offering a 95% discount on the input-token rate for cached input tokens. This pricing structure is intended for the introductory period only. After this time, the costs are scheduled to increase to $4 per million input tokens and $20 per million output tokens.

The model is positioned as a professional tool rather than a general-purpose assistant for everyday users. It is specifically aimed at those handling demanding technical work in the software industry. This distinction helps clarify that the model's primary purpose is to assist in complex coding and security infrastructure.

The path toward wider release

The rollout of Gemini 4 Carbon represents a shift toward specialized AI models that can perform deep work. By focusing on the intersection of coding and security, Google is targeting a high-value niche in the enterprise market. The model's ability to work on a single problem for days sets it apart from typical generative AI tools.

As the model moves from a small group of defenders to a broader audience, the focus will likely remain on reliability and safety. The data gathered from the initial users will be critical in determining how the model handles diverse software environments. This feedback loop is essential for building trust with developers who rely on these tools for critical infrastructure.

Future updates will likely continue to refine the model's ability to generate secure code and identify exploits. As the technology matures, it may become a standard part of the software development lifecycle. For now, the focus remains on providing a powerful, specialized tool for the most demanding technical challenges in cybersecurity.

Google launches Gemini 4 Carbon for cybersecurity and coding

key_number_label_placeholder_used_in_json_structure_only